一、Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey
用于分类和目标检测的少样本类增量学习:综述 引用: @ARTICLE{10840313, author={Zhang, Jinghua and Liu, Li and Silvén, Olli and Pietikäinen, Matti and Hu, Dewen}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, title={Few-Shot Class-Incremental Learning for Classification and Object Detection: A Survey}, year={2025}, volume={47}, number={4}, pages={2924-2945}, keywords={Power capacitors;Surveys;Training;Systematics;Measurement;Object detection;Data privacy;Data models;Artificial neural networks;Overfitting;Incremental learning;continual learning;lifelong learning;class-incremental learning;catastrophic forgetting;few-shot learning;few-shot class-incremental learning;deep learning;image classification}, doi={10.1109/TPAMI.2025.3529038}}
IEEE Transactions on Pattern Analysis and Machine Intelligence ( IF 20.8 )
Pub Date : 2025-01-14 , DOI: 10.1109/tpami.2025.3529038
摘要:
少样本类增量学习(FSCIL)是机器学习(ML)领域的一项独特挑战,它要求模型能够基于**稀疏标注的训练样本**持续学习新类别,同时避免遗忘先前学到的知识。尽管该领域近年来取得了一定进展,但仍处于积极探索阶段。
本文旨在对FSCIL进行全面系统的梳理。通过深度剖析,我们重点探讨了以下维度:
1. 核心问题定义
2. 关键挑战分析:不可靠经验风险最小化与稳定性-可塑性困境
3. 通用解决方案框架
4. 相关研究脉络:增量学习(IL)与少样本学习(FSL)的关联性研究
研究同时涵盖了:
- 基准数据集与评估指标分析
- 少样本类增量分类(FSCIC)方法三大流派:
▶